mmuokubamichael/4th-place-solution-Mozilla-Luganda-Automatic-Speech-Recognition

# 4th place solution Mozilla Luganda Automatic Speech Recognition Voice recognition technologies can make services more accessible and more human. But the problem is that automatic speech recognition (ASR) requires enormous amounts of voice data. There have been many advances in data collection and Automatic Speech Recognition (ASR) methodologies in recent years, which are revolutionizing the possibilities in this area especially for under-resourced languages in Africa. The objective of this challenge is to build a machine learning ASR model to accurately recognize words spoken in Luganda. You are encouraged to use the resources available to you from NVIDIA, such as NVIDIA NeMo ASR and documentation on NVIDIA ASR. You can also use transfer learning using the Kinyarwanda language datasets available on Mozilla Common Voice, and any other open-source models and libraries. A more accurate ASR model for Luganda could lead to new mobile applications and other products and services that better serve Lugandan speakers. From transportation and healthcare to financial services and education, voice recognition for under-resourced languages has the potential to change the daily realities for millions of people. To be eligible to win the NVIDIA and GTC prizes, you must register for GTC before 31 October 2021. You can register here. The dataset has been made possible through the work of Makerere AI lab. The lab is working towards building a Luganda data set through the Common Voice platform as part of their work of building NLP text and speech data set for low resourced languages in East Africa. link to the competition: zindi.africa In this solution, A wav2vec transformer model from hugginface was fine turned with luganda mozilla dataset with the exception of a language model ## 🚀 About Me I'm a full stack developer and a machine learning engineer...